Watermarking as AI content detection remains theoretically feasible but operationally constrained by accuracy, governance, and open-source incompatibility

A Brookings analysis of AI content detection methods—particularly watermarking—concludes that while statistical and machine learning-based watermarks can raise barriers to evasion, no single technical approach satisfies all policy and technical requirements simultaneously. Watermarks require model-developer cooperation, are undermined by open-source release, and face forgery risk; post-hoc detectors are universal but accuracy-challenged. The report proposes a realistic objective: raising evasion barriers high enough to identify a major fraction of AI-generated content, particularly from popular closed models, while acknowledging watermarking will be of limited use in high-stakes settings. For financial institutions, this implies regulatory frameworks requiring AI content labeling face implementation constraints that may delay widespread detection infrastructure and create ongoing compliance uncertainty.